Building a Language Learning Game Taught Me Something Unexpected About AI

When we started experimenting with AI translations, we assumed the biggest challenge would be accuracy. We were wrong. The harder problem was preference. Give two AI models the same sentence, and both translations can be technically correct. Yet people almost always have a favorite. One sounds natural. One feels human. One is the version they'd actually use. That observation eventually led us to build Parley, a simple game where players compare two translations and choose the better one. What happened next surprised us. People became highly engaged with a task that looked almost trivial.